| ---
|
| language:
|
| - en
|
| - vi
|
| license: other
|
| library_name: gguf
|
| tags:
|
| - acne
|
| - dermatology
|
| - skincare
|
| - gguf
|
| - qwen3.5
|
| - bilingual
|
| pipeline_tag: text-generation
|
| base_model:
|
| - Qwen/Qwen3.5-0.8B
|
| ---
|
|
|
| # Acnoryx AI Release
|
|
|
| ## Overview
|
|
|
| - Model family: Qwen/Qwen3.5-0.8B
|
| - Project name: acnoryx
|
| - Model size: 0.8b
|
| - GGUF quantizations: F16, Q8_0, Q5_K_M, Q4_K_M, Q4_0, IQ4_NL, IQ4_XS
|
| - Domain: acne, acne-prone skin, skincare, and dermatology guidance
|
|
|
| ## App
|
|
|
| - Google Play: https://play.google.com/store/apps/details?id=com.fivecanh.acnoryx
|
|
|
| ## Default behavior
|
|
|
| - The saved tokenizer/chat template injects a short default system prompt when no system message is provided.
|
| - That means the model can still understand its identity as Acnoryx AI in chat mode without a long system prompt.
|
| - For best results, write Vietnamese with full accents, or use natural English.
|
|
|
| ## Prompt examples
|
|
|
| - Tiếng Việt: `Da em nhiều mụn viêm ở má, routine hiện tại chỉ có sữa rửa mặt và kem dưỡng. Em nên ưu tiên gì trước?`
|
| - Tiếng Việt: `Kết quả quét của tôi có mụn đầu đen 32%, mụn mủ 21%, thâm mụn 18%. Hãy tóm tắt đúng theo dữ liệu.`
|
| - English: `I have oily acne-prone skin with dark marks after breakouts. What should I prioritize first?`
|
|
|
| ## Included folders
|
|
|
| - gguf/: GGUF exports for llama.cpp runtimes
|
| - hf_transformers/: merged Hugging Face Transformers model
|
|
|
| ## Training stack
|
|
|
| - Transformers + PEFT + TRL bf16 LoRA
|
| - Qwen3.5 hybrid architecture with fast linear path enabled when available
|
|
|
| ## Prompting
|
|
|
| - See PROMPT_TEMPLATE.txt for usage guidance.
|
|
|
| ## Evaluation Snapshot
|
|
|
| Release GGUFs were retested on the curated `release_eval_v1` set with 58 bilingual questions in both thinking and non-thinking modes.
|
|
|
| | Quant | Think | No-Think | Avg | Notes |
|
| |---|---:|---:|---:|---|
|
| | Q8_0 | 86.2% | 87.9% | 87.0% | Best overall score in the current release rerun |
|
| | Q5_K_M | 89.7% | 82.8% | 86.2% | Strong think-mode quality |
|
| | IQ4_NL | 86.2% | 86.2% | 86.2% | Best balanced sub-500 MB option |
|
| | F16 | 87.9% | 81.0% | 84.4% | Highest-fidelity source export |
|
| | IQ4_XS | 84.5% | 81.0% | 82.8% | Smaller release option |
|
| | Q4_K_M | 82.8% | 81.0% | 81.9% | Usable but clearly weaker than Q8_0 / Q5_K_M |
|
| | Q4_0 | 77.6% | 75.9% | 76.8% | Lowest-quality release quant |
|
|
|
| ## Deployment Guidance
|
|
|
| - Recommended default release quant: **Q8_0**
|
| - Best size/quality trade-off under 500 MB: **IQ4_NL**
|
| - Keep **Q4_0** only for constrained experiments, not as a primary deployment target
|
| - Current release family remains below the older internal 96% gate, so these artifacts should be treated as interim bundles rather than a final quality-signoff build
|
|
|
| ## Test Results
|
|
|
| Latest automated GGUF test results are below:
|
|
|
| - **acnoryx-0.8b-f16**: Think mode 87.9%, No-Think mode 81.0%
|
| - **acnoryx-0.8b-iq4_nl**: Think mode 86.2%, No-Think mode 86.2%
|
| - **acnoryx-0.8b-iq4_xs**: Think mode 84.5%, No-Think mode 81.0%
|
| - **acnoryx-0.8b-q4_0**: Think mode 77.6%, No-Think mode 75.9%
|
| - **acnoryx-0.8b-q4_k_m**: Think mode 82.8%, No-Think mode 81.0%
|
| - **acnoryx-0.8b-q5_k_m**: Think mode 89.7%, No-Think mode 82.8%
|
| - **acnoryx-0.8b-q8_0**: Think mode 86.2%, No-Think mode 87.9%
|
|
|
| Full detailed results in `results/release_gguf_0.8b/TEST_RESULTS.json`.
|
|
|
| For cross-family comparison with research quants, see `results/COMPARISON.md` in the workspace.
|
|
|
